Tachintech/TachinGlove_EMF

TachinGlove EMF

Visualization Preview

The default Dataset Viewer configuration contains representative screen recordings from the included Rerun visualization tool. Each video presents synchronized stereo EGO views, EMF hand skeletons and fingertip poses, tactile visualization, and EGO IMU signals. These preview videos are derived visualizations for convenient inspection; the original sensor files remain the authoritative dataset records.

TachinGlove EMF is an embodiment-free multimodal dataset sample collected with TachinGlove Pro-E tactile data gloves and a stereo EGO camera system. It provides synchronized human-hand demonstrations that are not tied to a particular robot embodiment.

TachinGlove Pro-E integrates tactile sensing and electromagnetic-field (EMF) tracking. The system captures hand poses together with tactile signals from the palm and fingers. The accompanying stereo EGO system records left- and right-view video, while an EGO IMU records inertial measurements.

The dataset focuses on tactile-rich dexterous manipulation. It currently contains 369 recording sessions across five provisional task categories. The labels and taxonomy are still being organized and may be refined in later releases.

Status: This is an initial example release. Field-level specifications, sensor units, effective sampling rates, timestamp synchronization rules, and the complete EMF coordinate-frame convention are still being documented. The per-session meta_data.json files should be treated as the most authoritative machine-readable record for the current release.

Intended Uses

Potential research uses include:

• tactile-rich dexterous manipulation;
• hand-pose and fingertip-motion analysis;
• tactile perception and contact modeling;
• multimodal sensor fusion;
• learning from human demonstrations;
• embodiment transfer and robot learning;
• synchronization and visualization of tactile, EMF, video, and IMU streams.

This release is intended for non-commercial research and education. See License for details.

Data Collection Hardware

TachinGlove Pro-E

TachinGlove Pro-E is a tactile and EMF data-collection glove. In this dataset, a left and a right glove record:

• raw tactile measurements from distributed tactile patches;
• EMF-based five-fingertip poses;
• EMF-based 25-joint hand skeletons;
• timestamps and session-level metadata.

Stereo EGO system

The stereo EGO system records:

• raw left-camera video;
• raw right-camera video;
• per-frame hardware and host-computer timestamps;
• EGO IMU gyroscope and accelerometer streams.

Task Categories

The current directory labels are provisional and remain under active organization.

Folder label	Description	Sessions
Button & Knob Manipulation	Interactions with buttons, knobs, and related controls	56
Fabric Grasping & Flattening	Grasping, repositioning, and flattening fabric	20
Open & Close	General opening and closing operations	120
Plug & Unplug	Plugging and unplugging objects or connectors	150
Twist Open & Close	Twist-based opening and closing, such as bottles or cups	23
Total		369

The task category is encoded by the parent folder name. Individual session folders currently use automatically generated acquisition identifiers rather than semantic clip names.

Repository Structure

TachinGlove_EMF/
├── Sample_data_EMF_1023/
│   ├── Button & Knob Manipulation/
│   │   └── customized_glove_device_pack_id_0_<timestamp>/
│   ├── Fabric Grasping & Flattening/
│   │   └── customized_glove_device_pack_id_0_<timestamp>/
│   ├── Open & Close/
│   │   └── customized_glove_device_pack_id_0_<timestamp>/
│   ├── Plug & Unplug/
│   │   └── customized_glove_device_pack_id_0_<timestamp>/
│   └── Twist Open & Close/
│       └── customized_glove_device_pack_id_0_<timestamp>/
└── TachinGlove_EMF_Pointclouds_Pack/
    └── emf_pointclouds_pack/
        ├── host_session_viewer.py
        ├── tactile_skin.py
        ├── tactile_vis.py
        ├── glove_seg_params.json
        ├── requirements.txt
        ├── README.md
        ├── README_EN.md
        └── tactile_skin/

Session Contents

A typical session contains the following files:

customized_glove_device_pack_id_0_<timestamp>/
├── meta_data.json
├── ego_imu.csv
├── video_left_cam0.mp4
├── video_right_cam1.mp4
├── video_left_ts.csv
├── video_right_ts.csv
├── left_hand_data.npz
├── right_hand_data.npz
├── emf_fingertips_left_<timestamp>.npz
├── emf_fingertips_right_<timestamp>.npz
├── emf_skeleton_left_<timestamp>.npz
└── emf_skeleton_right_<timestamp>.npz

File availability and frame counts should be checked against the session's meta_data.json. Consumers should not assume that every stream has identical length or sampling frequency.

File Summary

File pattern	Modality	Current contents
meta_data.json	Metadata	Device/session identifiers, stream status, schema versions, clock information, frame counts, completeness flags, errors, and acquisition metadata
ego_imu.csv	EGO IMU	Timestamp, three-axis gyroscope, and three-axis accelerometer measurements
video_left_cam0.mp4	Video	Raw left EGO camera stream
video_right_cam1.mp4	Video	Raw right EGO camera stream
video_left_ts.csv	Video timestamps	Frame indices plus left-camera hardware and host timestamps
video_right_ts.csv	Video timestamps	Frame indices plus right-camera hardware and host timestamps
left_hand_data.npz	Tactile	Raw left-glove tactile arrays, frame numbers, timestamps, and per-patch transform maps
right_hand_data.npz	Tactile	Raw right-glove tactile arrays, frame numbers, timestamps, and per-patch transform maps
emf_fingertips_left_*.npz	EMF fingertips	Left fingertip positions, orientations, validity flags, and timestamps
emf_fingertips_right_*.npz	EMF fingertips	Right fingertip positions, orientations, validity flags, and timestamps
emf_skeleton_left_*.npz	EMF skeleton	Left 25-joint positions, orientations, coordinate-frame label, and timestamps
emf_skeleton_right_*.npz	EMF skeleton	Right 25-joint positions, orientations, coordinate-frame label, and timestamps

Observed Data Schemas

The following names were observed in a representative session from this release. They document the current files but do not yet constitute a frozen public schema.

EGO IMU CSV

Observed header:

timestamp_ns,gyro_x,gyro_y,gyro_z,accel_x,accel_y,accel_z

Field	Description	Unit/status
timestamp_ns	IMU sample timestamp	Nanoseconds, based on the field name; full clock definition pending
gyro_x, gyro_y, gyro_z	Three-axis gyroscope readings	To be documented
accel_x, accel_y, accel_z	Three-axis accelerometer readings	To be documented

Video timestamp CSV

Observed header for both left and right streams:

file_seq,frame_index,hw_timestamp_s,pc_timestamp_s

Field	Description	Unit/status
file_seq	Video file sequence identifier	Integer
frame_index	Frame index within the recording	Integer
hw_timestamp_s	Camera hardware timestamp	Seconds; epoch/clock semantics to be documented
pc_timestamp_s	Host-computer timestamp	Seconds; synchronization semantics to be documented

EMF fingertip NPZ

Observed arrays:

received_timestamp_sec
monotonic_timestamp_sec
positions_mm
quaternions_xyzw
valid
metadata_json

The five fingertips are ordered as recorded in meta_data.json; the observed order is thumb, index, middle, ring, and pinky. Positions are labeled in millimeters and quaternions use xyzw ordering. Fingertip data is labeled in the emf_reference frame.

EMF skeleton NPZ

Observed arrays:

host_time_ns
host_write_ns
host_monotonic_ns
sdk_timestamp_sec
emf_received_timestamp_sec
emf_monotonic_timestamp_sec
coordinate_frame
positions_local_mm
orientations_local_xyzw
metadata_json

The observed skeleton contains 25 joints:

wrist
thumb_metacarpal, thumb_proximal, thumb_distal, thumb_tip
index_metacarpal, index_proximal, index_intermediate, index_distal, index_tip
middle_metacarpal, middle_proximal, middle_intermediate, middle_distal, middle_tip
ring_metacarpal, ring_proximal, ring_intermediate, ring_distal, ring_tip
little_metacarpal, little_proximal, little_intermediate, little_distal, little_tip

Positions are labeled in millimeters, orientations use xyzw quaternion ordering, and the current skeleton data is labeled in the wrist_local frame. The observed EMF files use schema_version=5.

Tactile NPZ

Observed arrays include:

timestamps
frame_numbers
tactile_<patch_id>
tf_tactile_x_<patch_id>
tf_tactile_y_<patch_id>

Observed tactile patch IDs are:

0, 1, 2, 3, 4, 5, 7, 8, 9, 11, 12, 13, 15, 16, 18

The numeric patch IDs are legacy acquisition identifiers and are intentionally preserved because they are part of the NPZ array names. For documentation and downstream processing, the following semantic aliases are recommended:

Sensor key	Recommended semantic alias	Anatomical region (illustrative)
tactile_0	thumb_distal_pad	Distal thumb / thumb-tip pad
tactile_1	thumb_middle_pad	Middle thumb pad
tactile_2	thumb_thenar_pad	Thumb base / thenar region
tactile_3	index_distal_pad	Distal index-finger pad
tactile_4	index_middle_pad	Middle index-finger pad
tactile_5	index_proximal_pad	Proximal index-finger pad
tactile_7	middle_finger_distal_pad	Distal middle-finger pad
tactile_8	middle_finger_middle_pad	Mid-segment middle-finger pad
tactile_9	middle_finger_proximal_pad	Proximal middle-finger pad
tactile_11	ring_distal_pad	Distal ring-finger pad
tactile_12	ring_middle_pad	Middle ring-finger pad
tactile_13	ring_proximal_pad	Proximal ring-finger pad
tactile_15	little_distal_pad	Distal little-finger pad
tactile_16	little_middle_pad	Middle little-finger pad
tactile_18	palm_hypothenar_pad	Palm / hypothenar region

This mapping describes the right-hand glove in palmar view; the left-hand glove uses the mirrored layout. The release contains 15 tactile pads. The numeric IDs are non-contiguous by design in the current acquisition format and should not be interpreted as spatial indices or anatomical ordering.

These aliases are documentation-level names only: the released NPZ keys remain tactile_<patch_id>, and downstream code should retain the numeric ID when reading the current data. Exact patch geometry and taxel locations are defined by the left- and right-hand tactile-skin schema files included with the visualization toolkit.

The physical units, value range, calibration procedure, taxel geometry, and complete meanings of the tf_tactile_* arrays are to be documented. The included visualization package contains tactile-skin schemas and calibration parameters used for rendering.

Session metadata JSON

meta_data.json contains session-level information such as:

• device and session identifiers;
• left/right glove sources;
• EMF schema and algorithm versions;
• position, time, and quaternion metadata;
• EMF node and fingertip names;
• per-stream frame counts;
• recording start and stop timestamps;
• completion flags and recorded errors;
• tactile output files and frame counts;
• video start and hardware-start timestamps.

Some metadata values may describe acquisition internals rather than stable public API fields.

Units, Sampling Rates, and Synchronization

The complete specification is still being prepared.

Current files explicitly encode several units in field or metadata names, including positions_mm, timestamps ending in _sec or _s, and host timestamps ending in _ns. However, users should consult each session's meta_data.json and avoid assuming that all modalities share one clock or sampling frequency.

The following items remain to be documented:

• tactile signal units, range, and calibration;
• gyroscope and accelerometer units;
• nominal and effective sampling rate for each modality;
• dropped-frame and invalid-frame handling;
• timestamp epoch and clock domain for every timestamp field;
• synchronization accuracy and alignment procedure across gloves, cameras, and IMU;
• recommended interpolation or resampling procedure.

EMF Coordinate Frames

The current release contains two distinct EMF coordinate frames. They must not be treated as the same physical frame.

EMF skeleton: wrist_local

The emf_skeleton stream stores a 25-joint hand skeleton in the wrist_local coordinate frame.

• The wrist is the local origin.
• Position arrays are stored as positions_local_mm and are labeled in millimeters.
• Orientation arrays are stored as orientations_local_xyzw and use xyzw quaternion component ordering.
• The observed files use schema_version=5.
• The session metadata records skeleton_motion=wrist_local_only.

Schema v5 does not retain the global/root wrist position or root wrist quaternion. Consequently, the saved skeleton cannot be transformed back into the EMF base-station frame from the released skeleton file alone.

EMF fingertips: emf_reference

The emf_fingertips stream stores the poses of five fingertip sensors in the emf_reference coordinate frame.

• The origin is the EMF reference/base station mounted on the back of the hand.
• The origin is not the wrist center and is not the origin of the skeleton's wrist_local frame.
• Position arrays are stored as positions_mm and are labeled in millimeters.
• Orientation arrays are stored as quaternions_xyzw and use xyzw quaternion component ordering.
• Fingertips are ordered as thumb, index, middle, ring, and pinky, as recorded in meta_data.json.
• The valid array indicates the validity of each fingertip measurement.

The current release does not contain a measured rigid transformation from the back-of-hand emf_reference frame to the skeleton's wrist_local frame. Recovering this physical transformation would require the root wrist position and orientation in the base-station frame to be retained during acquisition.

Visualization-only axis mapping

The included viewer applies an axis permutation to the fingertip data before displaying it with the wrist-centered skeleton. This mapping is a display convention, not a measured base-station-to-wrist extrinsic calibration.

For column-vector positions, the viewer applies:

p_mapped = M @ p_emf_reference

The matrices are:

Hand	Mapping matrix M	Component form
Right	[[0, 0, -1], [0, 1, 0], [1, 0, 0]]	(x, y, z) -> (-z, y, x)
Left	[[0, 0, 1], [0, -1, 0], [1, 0, 0]]	(x, y, z) -> (z, -y, x)

After this axis mapping, the visualization applies the wrist-derived display rotation R_full and the same side-by-side display translation side_offset used for the wrist center:

R_mapped = M @ R_emf_fingertips
position_display = R_full.T @ M @ position_emf + side_offset
orientation_display = R_full.T @ R_mapped

For the skeleton display, R_full is estimated from the wrist, fingertips, and palm width so that the palm faces approximately +Z and the fingers point approximately toward -Y. Additional display corrections are applied so that the two hands use visually consistent local axes:

• Right-hand whole-palm correction: diag(-1, 1, -1).
• Left-hand local joint correction: diag(-1, -1, 1) after applying R_full.T @ R_emf_skeleton.
• Left-hand *_tip nodes receive the same local correction once more because their saved local orientation already contains the corresponding 180-degree rotation relative to the distal segment.

The viewer places the emf_reference origin at the displayed wrist location by reusing the wrist display translation. This is only a visualization choice: fingertip points are not physically recentered onto the wrist and are not snapped to the skeleton's *_tip joints.

Information not yet specified

The following native EMF conventions remain to be documented:

• the physical directions of the native emf_reference +X, +Y, and +Z axes;
• whether each native frame is defined as right-handed or left-handed;
• whether stored quaternions represent active or passive rotations;
• whether each quaternion maps local-to-reference or reference-to-local coordinates;
• whether the native left- and right-hand frame conventions are identical;
• the measured rigid extrinsic transformation between the back-of-hand base station and the wrist center;
• extrinsic transformations between the EMF system, stereo EGO cameras, and EGO IMU.

Visualization Toolkit

The repository includes a Rerun-based visualization package:

TachinGlove_EMF_Pointclouds_Pack/emf_pointclouds_pack

It can display:

• left and right 25-joint EMF hand skeletons;
• five EMF fingertip poses per hand;
• tactile-patch point clouds attached to hand joints;
• optional stereo MP4 video;
• optional tactile heatmaps;
• optional EGO IMU streams;
• interactive tactile-patch calibration controls.

Requirements

• Python 3.10 or later
• dependencies listed in requirements.txt, including NumPy, OpenCV, Pillow, and the Rerun SDK

Installation

cd TachinGlove_EMF_Pointclouds_Pack/emf_pointclouds_pack
python -m pip install -r requirements.txt

Run a session

python host_session_viewer.py --session-dir "/absolute/path/to/a/session"

For example, after downloading the repository:

python host_session_viewer.py \
  --session-dir "../../Sample_data_EMF_1023/Plug & Unplug/customized_glove_device_pack_id_0_20261022_143248_743"

List available sessions under a data root:

python host_session_viewer.py \
  --list \
  --data-root "../../Sample_data_EMF_1023"

Run a faster preview with reduced video/IMU sampling and without video images:

python host_session_viewer.py \
  --session-dir "/absolute/path/to/a/session" \
  --fast \
  --no-images

Run the interactive tactile-patch calibration interface:

python host_session_viewer.py \
  --tune-glove \
  --session-dir "/absolute/path/to/a/session"

See the toolkit documentation for coordinate transforms, tactile-patch attachment, calibration controls, and additional options:

• README_EN.md
• README.md

Download

Using the Hugging Face CLI:

hf download Tachintech/TachinGlove_EMF \
  --repo-type dataset \
  --local-dir TachinGlove_EMF

The repository contains large binary files. Make sure sufficient disk space is available before downloading the full dataset.

Dataset Limitations

• This is an example release rather than a finalized benchmark.
• Task labels and the task taxonomy are still being organized.
• Session folder names are acquisition identifiers rather than semantic descriptions.
• Some streams may have different lengths, clock domains, and effective sampling rates.
• Complete field-level, synchronization, and coordinate-frame specifications are not yet available.
• EMF measurements may be affected by electromagnetic interference and local acquisition conditions.
• Tactile measurements may depend on glove fit, contact geometry, sensor calibration, and the wearer.
• The dataset is not designed to represent all users, hand sizes, objects, environments, or manipulation strategies.
• The supplied visualization transforms are intended for display and should not automatically be treated as metric extrinsic calibration.

Privacy and Ethics

This dataset may contain egocentric video of the data-collection environment. Users should avoid attempting to identify participants or infer sensitive information from the recordings. Any redistribution or derivative release must follow the dataset license and applicable privacy, consent, and research-ethics requirements.

Additional documentation about participant consent, anonymization, and the recording environment will be added as the release is finalized.

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

It is free to use for non-commercial research and education, provided appropriate attribution is given.

For commercial use, please contact TachinTech to obtain separate permission.

Citation

If you use this dataset, please cite the dataset repository:

@dataset{tachinglove_emf_2026,
  title     = {TachinGlove EMF: An Embodiment-Free Multimodal Dataset Sample for Tactile-Rich Dexterous Manipulation},
  author    = {{TachinTech}},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Tachintech/TachinGlove_EMF}
}

Contact

For questions, corrections, commercial licensing, or collaboration, please contact TachinTech through the repository's Community tab.